GPU Passthrough in Proxmox: Complete Guide for Homelabs

So you’ve got a GPU sitting idle while your VMs run on CPU alone. Or maybe you’re tired of maintaining a separate machine just for gaming or AI workloads. GPU passthrough is the answer—and Proxmox makes it surprisingly approachable.

Let’s walk through everything you need to know: what it is, whether your hardware can handle it, and how to set it up without pulling your hair out.

What Is GPU Passthrough?

GPU passthrough gives a virtual machine direct access to a physical graphics card. Instead of virtualizing the GPU (which comes with overhead), the VM gets near-native performance because it talks directly to the hardware.

Think of it like handing someone the keys to your car versus driving them around in a sidecar. One is way more performant.

Why Bother?

AI and Machine Learning. This is the big one right now. Local LLM inference with Ollama, training models with PyTorch, running CUDA workloads—all of these need a real GPU. Consumer cards work great for homelab setups.

Gaming in a VM. Run Windows in a VM with full GPU performance for gaming, then pause or snapshot it when you’re done. No dual-boot hassle.

Media Transcoding. Plex and Jellyfin hardware transcoding (NVENC for NVIDIA, VAAPI for AMD) can handle multiple streams effortlessly when given direct GPU access.

Professional Workloads. 3D rendering, video editing, CAD—all benefit from GPU acceleration in the VM.

The Catch

A passed-through GPU is exclusive to that VM. The host can’t use it, and neither can other VMs. You’ll need a separate GPU (or iGPU) for the Proxmox host itself.

Hardware Requirements: Can You Even Do This?

Not every CPU and motherboard supports GPU passthrough. Here’s what you need.

CPU Support

Vendor	Feature Required	Supported Generations
Intel	VT-d	Haswell (4th gen) and newer
AMD	AMD-Vi (IOMMU)	Bulldozer and newer

Most modern consumer CPUs support this. Older AMD K10 chips need specific motherboards (890FX or 990FX chipset).

Quick check: Intel’s ARK database lets you filter by VT-d support. AMD has similar documentation.

Motherboard Requirements

IOMMU/VT-d enabled in BIOS — This is non-negotiable. Set it to Enabled, not Auto.
ACS (Access Control Services) — Required for proper device isolation. Server boards tend to have better support than consumer boards.
Above 4G Decoding — Required for modern GPUs with large VRAM.

GPU Compatibility

NVIDIA: Everything from GeForce 7 through RTX 40-series works. Consumer cards may need driver tweaks (we’ll cover Error 43 later), but they work. For multi-VM sharing, you’ll need Quadro, RTX A-series, or Tesla cards with vGPU licensing.

AMD: Radeon 5000-7000 series work. Navi cards (5000/6000/7000 series) have a reset bug—fixable, but annoying.

Intel: 5th-7th gen Core iGPUs support GVT-g (virtual GPU). Arc discrete GPUs can be passed through directly.

BIOS Settings Checklist

Before you start, make sure these are configured:

Setting	Value	Notes
VT-d / IOMMU	Enabled	Not Auto—force it on
Intel VT-x / AMD-V	Enabled	Hardware virtualization
Above 4G Decoding	Enabled	Required for 24GB+ VRAM GPUs
Resizable BAR / SAM	Disabled	Breaks passthrough in VMs
CSM / Legacy Boot	Disabled	Cleaner for single-GPU setups
Primary Display	iGPU or PCI	Use iGPU for host if available

Step-by-Step Proxmox Setup

Alright, let’s configure your Proxmox host for GPU passthrough.

1. Enable IOMMU in the Kernel

Edit /etc/default/grub and add the appropriate parameters to GRUB_CMDLINE_LINUX_DEFAULT.

For Intel:

GRUB_CMDLINE_LINUX_DEFAULT="quiet intel_iommu=on iommu=pt"

For AMD:

GRUB_CMDLINE_LINUX_DEFAULT="quiet amd_iommu=on iommu=pt"

The iommu=pt enables passthrough mode—bypasses hypervisor DMA translation for better performance.

Update GRUB:

update-grub

2. Load VFIO Modules

Add the VFIO modules to load at boot:

cat >> /etc/modules << EOF
vfio
vfio_iommu_type1
vfio_pci
EOF

Update initramfs:

update-initramfs -u -k all

3. Blacklist GPU Drivers

Prevent the host from claiming the GPU. Create /etc/modprobe.d/blacklist.conf:

For NVIDIA:

cat >> /etc/modprobe.d/blacklist.conf << EOF
blacklist nouveau
blacklist nvidia*
EOF

For AMD:

cat >> /etc/modprobe.d/blacklist.conf << EOF
blacklist amdgpu
blacklist radeon
EOF

4. Reboot and Verify

reboot

After reboot, verify everything loaded correctly:

# Check IOMMU is enabled
dmesg | grep -e DMAR -e IOMMU -e AMD-Vi

# Verify interrupt remapping (security feature)
dmesg | grep 'remapping'

# Confirm VFIO modules loaded
lsmod | grep vfio

# Check GPU is bound to vfio-pci
lspci -nnk | grep -A3 -i nvidia
# Should show: "Kernel driver in use: vfio-pci"

5. Configure the VM

You can do this through the GUI or CLI. Here’s both.

GUI Method:

Select your VM → Hardware → Add → PCI Device
Choose your GPU from the list
Enable “All Functions” if the GPU has audio on the same bus
Enable “PCI-Express” for q35 machines
Check “Primary GPU” if this is the main display

CLI Method:

# Basic passthrough (replace VMID with your VM number)
qm set 100 -hostpci0 01:00,pcie=on,x-vga=on

# For GPUs with audio device on same bus
qm set 100 -hostpci0 01:00,pcie=on,x-vga=on,romfile=vbios.bin

Recommended VM Configuration:

Setting	Value	Why
Machine Type	q35	Required for PCIe passthrough
BIOS	OVMF (UEFI)	Modern GPUs need UEFI
CPU Type	host	Passes through CPU features
CPU Affinity	Pinned cores	Better performance (optional)

6. Install Guest Drivers

Boot your VM and install the appropriate drivers:

Windows: Download from NVIDIA or AMD’s website. For headless setups, enable Remote Desktop first.
Linux: Install nvidia-driver or amdgpu-install. For CUDA workloads, install the CUDA toolkit.

Single GPU vs. Multi-GPU Setups

Multi-GPU (Recommended)

The cleanest setup: one GPU for the host, one for the VM. Your iGPU can serve as the host GPU if you have one.

Install Proxmox normally
Pass through the dedicated GPU to the VM
Host uses iGPU for console
No complicated scripts needed

Single GPU (It’s Possible, But…)

Running Proxmox and your VM on the same GPU requires hook scripts to swap the device between them. When the VM starts, the GPU detaches from the host; when it stops, it reattaches.

If you must go this route:

Use iGPU for host display if your CPU has one
Add video=efifb:off to kernel parameters
Create hook scripts for GPU detachment/reattachment
For AMD cards, install vendor-reset (see troubleshooting below)

It works, but it’s finicky. Consider yourself warned.

Common Problems and Fixes

Error 43 on Windows (NVIDIA)

Symptom: Device Manager shows “Code 43” for the GPU. Black screen.

Solutions:

Cause	Fix
KVM visibility	Add `-cpu host,kvm=off` to VM args
Resizable BAR enabled	Disable in BIOS
Driver version mismatch (vGPU)	Match guest driver to host version
Mobile/vGPU variant detection	Add device-id spoofing to VM config

For args workaround, add to your VM config:

args: -cpu host,kvm=off

AMD Reset Bug

Symptom: GPU works on first VM boot, fails on subsequent boots until host reboot.

Affected Cards: Navi architecture (5000/6000/7000 series)

Fix: Install the vendor-reset kernel module:

# Install build dependencies
apt install pve-headers-$(uname -r) build-essential git

# Build and install vendor-reset
git clone https://github.com/gnif/vendor-reset.git
cd vendor-reset
make
make install

# Load module at boot
echo "vendor-reset" >> /etc/modules
update-initramfs -u -k all

reboot

BAR 3 / BAR 0 Errors

“BAR 3: can’t reserve”:

Add to kernel command line:

video=efifb:off

“BAR0 is 0M” / Code 12:

Large VRAM GPUs (24GB+) exceed OVMF’s memory window. Set phys-bits:

qm set 100 --cpu host,phys-bits=host,flags=+pdpe1gb

Black Screen in VM

Cause	Solution
GPU not primary	Add `x-vga=on` to hostpci config
No driver installed	Boot with VNC, install guest drivers first
Wrong BIOS mode	Use OVMF for UEFI-capable GPUs
Windows Fast Startup	Disable in Windows power settings

Disable Windows Fast Startup:

powercfg -h off

Or: Control Panel → Power Options → “Choose what power button does” → Uncheck “Turn on fast startup”

HDMI Audio Crackling

Enable MSI for GPU audio:

echo "options snd-hda-intel enable_msi=1" >> /etc/modprobe.d/snd-hda-intel.conf
update-initramfs -u -k all
reboot

“Unsafe Interrupts” Error

Older hardware may lack interrupt remapping. Enable the workaround (slight security trade-off):

echo "options vfio_iommu_type1 allow_unsafe_interrupts=1" > /etc/modprobe.d/iommu_unsafe_interrupts.conf

Performance Tuning

You’ve got passthrough working. Let’s squeeze out more performance.

Host-Side Tuning

CPU Pinning:

Dedicate specific cores to your VM for reduced latency:

qm set 100 --affinity 0-3

Hugepages:

For latency-sensitive workloads (AI training, gaming), use hugepages:

# Add to /etc/default/grub
GRUB_CMDLINE_LINUX_DEFAULT="... hugepagesz=1G hugepages=16"

# Or dynamically
echo 16 > /sys/kernel/mm/hugepages/hugepages-1048576/nr_hugepages

VM Configuration

Setting	Optimal Value	Notes
Machine	q35	Better PCIe support
BIOS	OVMF	Modern GPUs require UEFI
CPU Type	host	Passes through all CPU features
Disk	VirtIO SCSI single + IO thread	Fastest disk controller
Disk Cache	Write-back	Faster; no-cache is safer but slower
Network	VirtIO	Paravirtualized, minimal overhead

GPU-Specific Notes

NVIDIA: Consumer GeForce cards limit NVENC encoding sessions (3-5 simultaneous). Quadro and datacenter cards have no limit. For Plex/Jellyfin transcoding, this matters.

AMD: No encoding session limits, but VAAPI quality varies. ROCm compute only works reliably on Linux guests—Windows support is experimental.

Verify Performance

# Inside Linux VM
nvidia-smi -l 1  # Watch GPU utilization
glxgears         # Basic OpenGL test

# For AI workloads
# Test CUDA/ROCm with your model

Well-configured passthrough gets you 95%+ of bare-metal GPU performance.

When Passthrough Won’t Work: Alternatives

vGPU (Virtual GPU)

Split one GPU across multiple VMs. Requires NVIDIA licensing and supported cards (RTX A-series, Quadro RTX, Tesla).

# Enable SR-IOV virtual functions
systemctl enable --now [email protected]

# Add vGPU to VM
qm set 100 -hostpci0 01:00.4,mdev=nvidia-660

Good for homelabs that need GPU across multiple VMs, but licensing cost is real.

Intel GVT-g

Virtualize Intel’s iGPU. Works on 5th-7th gen Core processors.

# Add kernel parameter
i915.enable_gvt=1

# Create vGPU for VM
qm set 100 -hostpci0 00:02.0,mdev=i915-GVTg_V5_4

Depreciated on newer Intel CPUs—use SR-IOV for 8th gen+.

SR-IOV

Hardware-level virtualization for network cards and some GPUs. Intel Arc GPUs support this natively. NVIDIA requires vGPU licensing for SR-IOV use on Ampere+ cards.

Recommended Configurations

Gaming VM

# Host config
GRUB_CMDLINE_LINUX_DEFAULT="quiet intel_iommu=on iommu=pt"

# VM config (100.conf)
machine: q35
bios: ovmf
cpu: host
cores: 8
memory: 16384
hostpci0: 01:00,pcie=on,x-vga=on

Use dedicated GPU for gaming VM, iGPU for host.

AI/ML Workload VM

# VM config
machine: q35
bios: ovmf
cpu: host
cores: 16
memory: 65536
hostpci0: 01:00,pcie=on

# For NVIDIA: Install CUDA toolkit in guest
# For AMD: Install ROCm (Linux only)

Consider hugepages for large models. Pin CPU cores for training workloads.

Media Server (Plex/Jellyfin)

# VM config
machine: q35
bios: ovmf
cpu: host
cores: 4
memory: 8192
hostpci0: 01:00,pcie=on

# Enable NVENC in Plex/Jellyfin
# No x-vga needed for headless transcoding

NVENC handles multiple transcodes efficiently. Consumer NVIDIA cards: 3-5 sessions. Quadro/datacenter: unlimited.

Wrapping Up

GPU passthrough in Proxmox opens up serious capability: local AI workloads, gaming VMs, hardware transcoding—all on one box. The setup has a learning curve, but once configured, it’s rock solid.

Quick checklist before you start:

Verify VT-d/IOMMU support in your CPU and motherboard
Enable it in BIOS (not Auto—Enabled)
Disable Resizable BAR
Use q35 machine type with OVMF
Install vendor-reset if you have an AMD Navi card

When it works, you get near-native GPU performance in a VM. That’s the power of proper virtualization: one machine, many workloads, no compromise.

Found this guide helpful? I’d love to hear about your passthrough setup—drop a comment or ping me on Mastodon.